feat(cost): model streaming summary resources

[zzylol]

Why

At-rest query costing does not account for continuously arriving data, bootstrap work, active window state, retained state, or update multiplicity.

What

Introduce analytical evidence and CPU, peak-memory, and source-I/O formulas for incremental summary maintenance over one comparison horizon. The model uses the existing DataArrival and lifecycle vocabulary and fails closed when required physical evidence is absent.

How

• Separate initial/bootstrap inputs from arriving rows.
• Charge insert, merge, subtract, delete, and read operations at their physical multiplicities.
• Count shared physical state once by identity.
• Keep source scans, transient workspace, and retained state dimensionally separate.

Before this PR

The analytical estimator could price at-rest query DAGs but had no representation for update-driven summary state.

After this PR

A downstream evidence provider can describe one streaming summary implementation and obtain a resource-dimensional estimate without using structural node counts.

Stack

Base: #330. The next PR adds lifecycle comparison and ranking.

Module boundary

Streaming-specific inputs live under summary_maintenance_cost. The module adds bootstrap, arrival, active/retained-window, and summary-operation evidence on top of the shared analytical estimator over physical-operator evidence; it is not a second physical cost system.